Mechanism: A tool result arrives as static text, with no reliable interactivity, persistence, or control over how an agent presents it. Why now: Comet says its Opik users increasingly evaluate and debug agents from terminals where agents already work.
The article reports that Comet’s author built a text waterfall for trace data and tested it on ten traces, including latency spikes, retry storms, and cascading failures. It also reports that when an agent took a recovery path, it summarized rather than displaying the chart. A separate test found that a 222-column terminal was served a renderer defaulting to 80 columns, while agent-mediated output exposed color escape codes as text. These are observations from one team’s experiment, not a controlled usability study.
Evidence boundary: This is a first-person practitioner account about one product and renderer; it does not establish that all terminal agents behave this way or quantify user outcomes. First implication: Keep precise structured trace data available alongside human-oriented terminal visualizations. First watch condition: Watch whether agent-facing tools begin returning paired human-readable views and machine-readable evidence as a deliberate interface contract. Comet’s findings make terminal output an interface-design problem: hierarchy, clipping, color, and referenceability have to work in a page that can scroll away, while the model may interpret visual encodings imprecisely.